whylabs/langkit

Extracts specialized threat signals including jailbreak attempts, prompt injection attacks, hallucination detection, and refusal patterns alongside standard quality metrics. Built as a modular UDF layer that integrates directly with whylogs' schema system, enabling composable metric pipelines with configurable performance trade-offs (throughput ranges from 2K+ chats/sec with light metrics to sub-1 chat/sec with full analysis). Designed for production LLM monitoring workflows, with outputs visualizable in the WhyLabs observability platform or analyzed independently.

976 stars. No commits in the last 6 months.